When Andrea was studying math as an undergrad, she was required to take an arts class in order to graduate, and soon discovered that she loved poetry. She learned that the process of writing a poem was often similar to solving a complex math problem — just starting with one part, and then doing one more, and gradually the rest is revealed. She enjoyed it so much that her first machine learning project in graduate school was on poetry/sonnet generation.
Andrea wanted to blend technology and art in her career and Getty Images turned out to be the perfect place to combine her two interests. Getty Images curates and manages a huge library of images and videos that are used in editorial news, websites, social media, billboards and more. She started as a data scientist two years ago, and is now leading the AI/machine learning team to develop new tools to help clients more effectively use Getty’s creative assets. She explains that many of their creative clients come to the site and don’t have the language to describe the types of images they’re looking for. Andrea’s team is building models to help clients find images when they don’t have the words to articulate their vision. Another machine learning project they are working on is how to identify awkwardly posed “stocky” photos vs. the more realistic photos that clients are looking for.
She is currently doing more painting than poetry, and also sees parallels between painting and computer vision. She explains that when you start with a blank canvas, you have to think about shapes, lines, negative space and colors. It’s a similar process to how a machine comprehends pixels and the relationships between colors, contrast, shapes and textures.
Andrea created a special photo exhibit for the WiDS Stanford 2020 Conference that illuminates concerns about image manipulation while also posing provocative questions about gender and leadership. In the exhibit, she used machine learning (style GAN) to transform pictures of US presidents, ranging from George Washington to Donald Trump, into female versions of those presidents. The style GAN is a machine-learning model that can manipulate an image in different dimensions, in this case, from masculine to feminine.
The project was born out of conversations around Generative Adversarial Networks (GANs), synthetic image generation and concerns about the implications of deep fakes in politics and our culture. She wanted a way to expose that concern in a humorous way. She also saw this as an opportunity to re-imagine our history. What would the world be like today if our presidents had all been women?
She says the first response to the exhibit is usually laughter, but then it also sparks questions like: What would it have been like if females had founded the country? What wars would have happened or not happened? What would our constitution be like? How would capitalism have evolved? It catalyzes a conversation about the qualities of great leaders what leadership means through a female lens.